The r=2 Neural Law

Trained models from several families and developers share the same regularity.

Kind
Paper
Programme
FCN Compression
Published
27 June 2026
Status
Published

What we found

The regularity held closely in every model we measured.

What we also established

The law describes trained models rather than offering a shortcut: we established that knowing the regularity does not, on its own, make compression materially better.

Why it matters

A regularity shared by models from different families and developers points to training itself, rather than any one model, as its source.

The paper

The r = 2 Law of Mantissa-Bit Bias in Trained Neural Weights

Oliver Cooper, Forlais Group. Published 27 June 2026.

Read the paperPDF, 582 KBDOI 10.5281/zenodo.20961807CC BY 4.0

Cite this record

Forlais Research (2026). The r=2 Neural Law. Forlais Research Library, record FR-F-04.

@misc{fr-f-04,
  author = {{Forlais Research}},
  title = {{The r=2 Neural Law}},
  howpublished = {Forlais Research Library, record FR-F-04},
  year = {2026},
  note = {27 June 2026}
}